[mvapich-discuss] Binding processes to cores

[Jonathan Perkins]

Hello Michael, I think you may be confusing cpu binding with rank
placement.  Please see my replies inline.

On Fri, Oct 21, 2011 at 4:17 PM, Michael E. Thomadakis <miket at tamu.edu> wrote:
> Hello,
>
> 1) Besides the two options “bunch” and “scatter” or the explicit
> manipulation of the -hostfile, do we have any other way to fine tune the
> mapping of ranks to core or sockets ?

Other than bunch and scatter, you can explicitly provide the rank to
core mapping via MV2_CPU_MAPPING.  How to use this is discussed in
section 6.3 of our userguide.

> 2) It would be convenient to have a way to select nodes and ranks as for
> instance in "round-robin mapping of ranks to nodes" or "assign 2 ranks /
> node" or "spread evenly ranks to nodes", etc. I have not seen an "easy" to
> do regular-pattern spreading of ranks to nodes. Did I miss documentation on
> this ?

Currently there is no option with mpirun_rsh that allows you to do
round-robin placement of ranks without resulting in lower performance.
 If you're interested in a shorthand way to specify your hostfile
while preserving high performance you can use the following syntax in
your hostfile...

host : numprocs

Example:
node1:8
node2:8
node3:8
node4:8

The above hostfile will put the first 8 ranks on node1, the next 8 on
node2, and so on.  Hope this helps.

> thanks.....
>
> Michael
>
>
>
> On 10/21/11 14:34, Jonathan Perkins wrote:
>>
>> Hi, MVAPICH2 does this by default if you do not disable this during
>> configure.
>>
>> Please take a look at the following pointers for more information.
>>
>> Overview:
>>
>> http://mvapich.cse.ohio-state.edu/support/user_guide_mvapich2-1.7.html#x1-500006.3
>>
>> Enabling/disabling affinity:
>>
>> http://mvapich.cse.ohio-state.edu/support/user_guide_mvapich2-1.7.html#x1-15900011.20
>>
>> Specifying process binding:
>>
>> http://mvapich.cse.ohio-state.edu/support/user_guide_mvapich2-1.7.html#x1-15200011.13
>>
>> http://mvapich.cse.ohio-state.edu/support/user_guide_mvapich2-1.7.html#x1-15300011.14
>>
>> On Fri, Oct 21, 2011 at 3:18 PM, vaibhav dutt
>> <vaibhavsupersaiyan9 at gmail.com>  wrote:
>>>
>>> Hi,
>>> I am trying to bind ranks to specific cores so that there is no process
>>> migration from one
>>> core to another during the program execution. How can I do it in
>>> MVAPICH2?
>>> Thanks
>>> _______________________________________________
>>> mvapich-discuss mailing list
>>> mvapich-discuss at cse.ohio-state.edu
>>> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>>>
>>>
>>
>>
>
>
> --
> % -------------------------------------------------------------------- \
> % Michael E. Thomadakis, Ph.D.  Senior Lead Supercomputer Engineer/Res \
> % E-mail: miket AT tamu DOT edu                   Texas A&M University \
> % web:    http://alphamike.tamu.edu              Supercomputing Center \
> % Voice:  979-862-3931                    Teague Research Center, 104B \
> % FAX:    979-847-8643                  College Station, TX 77843, USA \
> % -------------------------------------------------------------------- \
>
>



-- 
Jonathan Perkins
http://www.cse.ohio-state.edu/~perkinjo